A cutting plane algorithm for the capacitated connected facility location problem

Noname manuscript No. (will be inserted by the editor) A cutting plane algorithm for the capacitated connected facility location problem Stefan Gollowitzer Bernard Gendron Ivana Ljubić Received: date / Accepted: date Abstract We consider a network design problem that arises in the costoptimal design of last mile telecommunication networks. It extends the Connected Facility Location problem by introducing capacities on the facilities and links of the networks. It combines aspects of the capacitated network design problem and the single-source capacitated facility location problem We refer to it as the Capacitated Connected Facility Location Problem. We develop a basic integer programming model based on single-commodity flows. Based on valid inequalities for the capacitated network design problem and the singlesource capacitated facility location problem we derive several (new) classes of valid inequalities for the Capacitated Connected Facility Location Problem including cut set inequalities, cover inequalities and combinations thereof. We use them in a branch-and-cut framework and show their applicability and efficacy on a set of real-world instances. Keywords Capacitated Network Design Facility Location Connected Facility Location Mixed Integer Programming Models Telecommunications S. Gollowitzer and I. Ljubić Department of Statistics and Operations Research, Faculty of Business, Economics, and Statistics, University of Vienna, Brünner Straße 72, 1210 Vienna, Austria Tel.: +43-1-4277-38663 Fax: +43-1-4277-38699 E-mail: {stefan.gollowitzer,ivana.ljubic}@univie.ac.at B. Gendron Interuniversity Research Centre on Enterprise Networks, Logistics and Transportation (CIRRELT), and Department of Computer Science and Operations Research, Université de Montréal, C.P. 6128, Succursale Centre-Ville, Montréal, Canada H3C 3J7E-mail: gendron@iro.umontreal.ca

2 Stefan Gollowitzer et al. 1 Introduction Given a set of customers, a set of potential facility locations and some interconnection nodes, the goal of the Connected Facility Location problem (ConFL) is to find the minimum-cost way of assigning each customer to exactly one open facility, and connecting the open facilities via a Steiner tree. The sum of costs for the Steiner tree, the facility opening costs and the assignment costs needs to be minimized. This problem models a network design problem that arises in the design of last mile telecommunication networks when the fiber to the curb (FTTC) deployment strategy is applied (see, e.g., [19]). Contrary to the fiber to the home strategy, where each customer, i.e., household, has its own fiber-optic uplink, in the FTTC strategy some of the existing copper wire infrastructure is used. More precisely, in an FTTC network, fiber optic cables run from a central office to a cabinet serving a neighborhood. End users connect to this cabinet using the existing copper connections. Expensive switching devices are installed in these cabinets. The usage of the last d meters of copper wire between the customer and a switching device may significantly reduce deployment costs while still enabling broadband connections of reasonable quality. In more detailed planning of FTTC networks, capacities of the links and of multiplexer devices are limited and this aspect was not captured by the ConFL variants studied in the literature so far. In this paper we consider a new capacitated variant of the ConFL problem, that we will refer to as the Capacitated Connected Facility Location Problem (CapConFL). In a typical application from telecommunications (see, e.g., Wassermann [35] for more details regarding input parameters), demands of customers are given as the number of twisted copper lines that are to be served at the respective customer location. Switching (or multiplexer) devices have both capacity and demand. Capacity is defined in terms of the number of twisted copper lines a device can serve. The demand of a switching device is defined as the number of fiber-optic uplinks required to connect the device to the central office (which is further connected to the backbone network). The number of uplinks is fixed for each device and independent of the number of customers that are finally assigned to it. The CapConFL consists of deciding on the location of switching devices, the assignment of customers to these devices and the routing of the uplinks from the switching devices to the central office, while minimizing the overall investment costs. 1.1 Problem definition More formally, CapConFL can be defined as follows. The input is a graph G = (V, E S A R ) with the set of nodes V partitioned into the set of customers (R), the set of potential facility locations (F ) and the set of potential Steiner nodes (V \ (F R)). A root node r V \ (F R) represents the connection to a higher order (e.g., backbone) network. The network G S = (V S, E S ), where

A cutting plane algorithm for the capacitated connected facility location problem 3 V S := V \ R and E S := {e = {i, j} E i, j V S } is called the core network. The assignment network G R = (F R, A R ) consists of directed arcs between potential facilities and customers, i.e., A R = {(i, k) i F, k R}. The following input parameters are associated to the network: Facility opening cost f i 0, capacity v i > 0 and demand d i > 0 for each i F. Arc cost c e 0 and capacity u e > 0 for each e E S. Assignment cost c ij 0 for each (i, j) A R. Customer demand b k > 0 for each k R. The goal is to find a subnetwork of G consisting of the set of open facilities F, the set of core edges E S and the set of assignment arcs A R such that: (P1) Each customer is assigned to exactly one open facility using arcs from A R. (P2) The sum of customers demands assigned to a facility i does not exceed its capacity v i. (P3) In the core subnetwork induced by E S, we can simultaneously route the flow from the root node to satisfy the demand of all open facilities, without violating the edge capacities. (P4) The sum of assignment, facility opening and edge costs, given by e E c S e + i F f i + (i,j) A c ij, is minimized. R Obviously, by setting capacities u e =, for all e E S and v i =, for all i F, we obtain the previously studied ConFL problem. Figure 1 illustrates solutions for ConFL and CapConFL. Squares and triangles denote facilities and customers, respectively. A black fill indicates that a facility is open. A diamond denotes the root node. Solid edges are in the core network, dotted edges represent the assignments. In the CapConFL the limited facility capacities require two additional open facilities and a different assignment of customers to facilities. The limited edge capacities require additional edges in the core network. Notice that ConFL combines the Steiner tree problem and the uncapacitated facility location problem. On the other hand, CapConFL combines the single-source capacitated network design problem with the single-source capacitated facility location problem. To see this, consider a feasible CapConFL instance whose core graph has a star topology. One can easily transform this input graph into an instance of the single-source capacitated facility location problem: the facility opening costs for each i F are now defined as c e + f i where c e is the corresponding adjacent edge, and the assignment graph remains unchanged. Similarly, a feasible CapConFL instance in which the assignment arcs are such that each customer is adjacent to exactly one facility can be reduced into an instance of the single-source capacitated network design problem.

4 Stefan Gollowitzer et al. (a) ConFL solution (b) CapConFL solution Fig. 1 Feasible solutions of the ConFL and CapConFL problem, respectively. 1.2 Literature review Since CapConFL has not been considered before, we provide a detailed literature overview of three closely related problems: connected facility location, capacitated network design and single-source capacitated facility location. 1.2.1 Connected Facility Location Early work on ConFL mainly includes approximation algorithms. ConFL can be approximated within a constant ratio and the currently best-known approximation ratio is provided by Eisenbrand et al. [14]. Recently, heuristic approaches have been proposed by Ljubić [28] and Bardossy and Raghavan [3]. Gollowitzer and Ljubić [19] present and compare several formulations for ConFL, both theoretically and computationally. Some of these results will be discussed and related to the CapConFL later on. Arulselvan et al. [2] consider a time-dependent variant of the ConFL and present a branch-and-cut approach based on cover, cut set cover and degree balance inequalities. Leitner and Raidl [27] propose a branch-and-cut-and-price approach for a variant of ConFL with capacities on facilities. Cutting planes are used to ensure paths between the root and open facilities, while column generation is used for selecting open facilities and assigning customers to them. 1.2.2 (Single-Source) Capacitated Network Design Problems (CNDP) In a typical CNDP setting, a network is given with a limited capacity available on each edge. A subset of edges of minimum cost needs to be installed in the network such that commodities with multiple origins and multiple destinations can be routed through the network without violating installed edge capacities. There exists a large body of work on the CNDP and related problems.

A cutting plane algorithm for the capacitated connected facility location problem 5 It includes exact methods based on Lagrangian relaxation or decomposition [16, 22, 34, 8, 26, 15], heuristic methods based on tabu search, neighborhood search, slope scaling, local branching and Lagrangian relaxation [10, 9, 17, 18, 11, 25, 32]. Recent developments comprise a theoretical study and comparison of Benders, metric and cut set inequalities [7] and a hybrid method combining mathematical programming and neighbourhood search techniques [20]. Finally, Chouman et al. [5] present a branch-and-cut approach that compares several families of valid inequalities for the CNDP. A generalization of the single-source CNDP is the Local Access Network Design problem (LAN). In this problem multiple copies of each edge are available. The Local Access Network Design problem was studied by Raghavan and Stanojević [31], Salman et al. [33] and Ljubić et al. [29]. 1.2.3 The Single-Source Capacitated Facility Location Problem (SSCFLP) Aardal et al. [1] and Deng and Simchi-Levi [12] proposed MIP models and studied the corresponding polyhedra of the SSCFLP and related problems. Holmberg et al. [21] present a branch-and-bound method based on a Lagrangean heuristic, Diaz and Fernández [13] develop a branch-and-price approach based on a decomposition of the SSCFLP and Contreras and Díaz [6] propose a scatter search heuristic. Ceselli et al. [4] give an exhaustive computational evaluation of branch-and-cut and branch-and-price approaches for a general class of facility location problems that includes the SSCFLP. 1.3 Contribution and outline In Section 2 we introduce a basic integer programming model for CapConFL and discuss the relation of CapConFL and the Connected Facility Location problem. In particular, we show that a domination result between two sets of valid inequalities for ConFL does not hold for CapConFL. In Section 3 we derive cover and extended cover inequalities for the various knapsack type constraints in our model. In addition, we provide two generalizations of recently proposed cut set cover inequalities and cover inequalities for single cut sets. Separation procedures for these valid inequalities are discussed in Section 4. In Section 5 we illustrate the effectiveness of the proposed model and the valid inequalities by computational experiments on a set of new, realistic benchmark instances based on real data. Conclusions are given in Section 6. 2 Mixed integer programming models In this section we introduce a first basic model for the CapConFL. It is based on models familiar in the context of the SSCFLP and the CNDP. We then strengthen this model using concepts known from the Connected Facility Location problem [19].

6 Stefan Gollowitzer et al. Since all demands of open facilities have to be routed from a single source node, it can be shown (see, e.g., [29]) that without loss of generality we can replace the undirected core network G S by a bidirected graph in which each edge e E S is replaced by two directed arcs, except for the edges adjacent to the root node, where it is sufficient to consider outgoing arcs from r. The set of arcs of the bidirected core network will be denoted by A S. Since the flow routed through an edge will always be routed in one of the two opposite directions, we define cost and capacities as c ij = c e and u ij = u e, respectively, for each e = {i, j} in E S. The union of core and assignment arcs is denoted by A = A S A R. For a set of customers J R we denote the set of facilities that can serve these customers by F (J) = k J F (k), where F (k) := {i F : (i, k) A R }. Likewise, for I F we denote by R(I) = i I R(i) where R(i) := {k R : (i, k) A R }. For W V we denote the set of ingoing arcs by δ (W ). 2.1 The basic MIP model In our models we will use the following binary decision variables: { 1, if arc (i, j) is installed x ij = (i, j) A 0, else { 1, if facility i is installed z i = i F 0, else In addition, continuous flow variables g ij indicate the total amount of flow between the root r and all open facilities in F routed through arc (i, j) A. The following model combines the single-commodity flow (SCF) formulation for the CNDP (see, e.g., [31, 33]) with a formulation for the SSCFLP (see, e.g., [21]): (SCF ) s.t. min f i z i c ij x ij + ij A i F d l z l g ji g ij = l F d lz l ji A S ij A S 0 i = l i = r else i V S (1a) 0 g ij u ij x ij (i, j) A S (1b) b k x ik v i z i i F (1c) k R(i) i F (k) x ik z i i F, k R(i) (1d) x ik = 1 k R (1e) x ij {0, 1} (i, j) A (1f)

A cutting plane algorithm for the capacitated connected facility location problem 7 z i {0, 1} i F (1g) Constraints (1c)-(1e) are the strong relaxation of the SSCFLP. The assignment constraints (1e) model property (P1) and constraints (1c)-(1d) ensure property (P2). In constraints (1a)-(1b) we use the single-commodity flow variables to ensure property (P3). This model is intuitive, but it provides weak lower bounds, due to the following facts: 1) big-m constraints (1b) are used to model the arc capacities, and 2) the connectivity between the root and the open facilities, rather than between the root and the customers, is required. The model is impractical to solve in a branch-and-bound framework, even for medium sized instances (see, e.g., Putz [30]). By using the following capacitated cut set inequalities to replace constraints (1a) and (1b), we can project out the flow variables from the previous model (see, e.g., Ljubić et al. [29]): u ij x ij d l z l W V S \ {r} (Cut SCF ) ij δ (W ) l F W The obtained model contains an exponential number of inequalities and provides the same lower bounds as the corresponding flow model. However, inequalities (Cut SCF ) can be strengthened as follows: min(u ij, d l )x ij d l z l W V S \ {r} ij δ (W ) l F W l F W 2.2 Relations to Connected Facility Location and cut set inequalities Gollowitzer and Ljubić [19] studied MIP formulations for ConFL and provided a complete hierarchy of several MIP formulations with respect to the quality of their LP-bounds. Among others, two cut set-based formulations for ConFL were described. The models differ in the way they require connectivity. In the first model, connectivity is ensured between the root and any open facility as follows: x ij z l W V S \ {r}, l W F (Cut Z ) ij δ (W ) These inequalities state that for each open facility the edges on at least one path between the root node and the respective facility need to be installed. Additional assignment constraints (1d) and (1e) are required between the facilities and the customers. The second model replaces constraints (Cut Z ) by the following cut set inequalities that ensure connectivity between the root and every customer: x ij 1 W V \ {r}, W R (Cut X ) ij δ (W )

8 Stefan Gollowitzer et al. For ConFL it was shown that the second model provides theoretically stronger lower bounds, but is computationally outperformed by the first model on the set of benchmark instances considered there. Both sets of inequalities, (Cut Z ) and (Cut X ) are also valid for CapConFL. It is interesting to mention that, unlike for ConFL, for which inequalities (Cut Z ) are implied by the model with (Cut X ) constraints, the two families of inequalities can be used complementary to each other for CapConFL: Lemma 1 Inequalities (Cut Z ) and (Cut X ) both strengthen the LP-relaxation of the basic model (SCF). However, the MIP models (SCF)+(Cut X ) and (SCF)+(Cut Z ) are incomparable w.r.t. the quality of their LP-bounds. Proof It is not difficult to see that inequalities (Cut Z ) and (Cut X ) both strengthen the LP-relaxation of (SCF). To see that (Cut Z ) inequalities are not implied by (SCF)+(Cut X ), consider the example shown in Figure 2. A vector (x, z) that satisfies (Cut SCF ) is x 12 = 0.75, x 23 = x 24 = 0.25, z 3 = z 4 = 0.75, x 35 = x 46 = 0.75 and x 45 = x 36 = 0.25. This solution is cut off by the (Cut X ) constraints x 23 + x 24 1 and x 12 1. Finally, inequalities (Cut Z ) are not redundant for (SCF)+(Cut X ) since they ensure x 23 + x 24 1.5 which further strengthens the model. Conversely, the model (SCF)+(Cut Z ) does not imply (Cut X ) constraints, which follows from the previous results for ConFL in [19], i.e., a CapConFL instance with sufficiently large capacities on arcs and facilities will have the desired property. u = 6 d 3 = 2 v 3 = 8 3 b 5 = 6 5 u = 4 1 2 u = 6 4 d 4 = 2 v 4 = 8 6 b 6 = 6 Fig. 2 Example for comparison of cut set inequalities 3 Valid Inequalities For the well-known subproblems of CapConFL, SSCFLP and CNDP, several sets of strengthening valid inequalities are known. We will review ideas that seem relevant in the context of the CapConFL and propose several sets of new valid inequalities based on the combination of the facility location and network design aspect.

A cutting plane algorithm for the capacitated connected facility location problem 9 3.1 Cover inequalities for single facilities Deng and Simchi-Levi [12] proposed cover inequalities for the SSCFLP with uniform capacities. These inequalities are better known in the context of general mixed integer programming to strengthen knapsack-type constraints. We will use the concept of extended cover inequalities (see, e.g., the recent work of Kaparis and Letchford [24]). Consider an arbitrary potential facility node i F. We call a set R R(i) a cover for i F if k R b k > v i and minimal if k R b k b l v i for all l R. For a minimal cover R, we define E(R ) = {k R(i) \ R : b k b }, where b = max k R b k. Let the set of all minimal covers of i F be denoted by MC(i). Then the following extended knapsack cover inequalities are valid for the CapConFL: x ij ( R 1)z i R MC(i), i F (EKS) j R E(R ) 3.2 Inequalities involving multiple facilities We derive two new families of inequalities that are implied by the limited capacities of facilities and the limited number of assignments edges in A R. 3.2.1 Minimum cardinality inequalities on facilities For a given set of customers J R and the corresponding subset of facilities F (J), let p(j) be the minimum number of facilities in F (J) that is required to assign the customers in J in a feasible way, i.e., by respecting the allowed possible assignments and satisfying the capacity constraints on the facilities in F (J). In other words, p(j) is the optimum solution of a capacitated binpacking problem with the set of bins F (J), capacities v i for i F (J), the set of items J, demands b j for j J and such that each item j J is only allowed to be assigned to bins in F (j). W.l.o.g. we can assume that b k v i for all (i, k) A R and thus p({k}) = 1 for all k R and p(j) min{ F (J), J } for all J R. Then the following minimum cardinality inequalities are valid for the Cap- ConFL: z i p(j) J R (MC F ) i F (J) 3.2.2 (Extended) Cover inequalities on facilities Next we apply the idea of cover inequalities to the relation of facility capacities and customer demands. Let again be J R. We call a set F F (J) a capacity cover with respect to J if i F (J)\F v i < b(j) and we call it minimal if v k + i F (J)\F v i b(j) for all k F. Let CC(F (J)) denote the set of all

10 Stefan Gollowitzer et al. such capacity covers of F (J). We call the following set of constraints cover inequalities on facilities: i F z i 1 F CC(F (J)), J R (2) Similar to the cover inequalities for single facilities we can extend the covers and obtain stronger inequalities. Let v = max i F v i and let E(F ) = {i F (J) \ F : v i v } be the set of remaining facilities from F (J) with a capacity of at least v. We refer to the following inequalities as extended cover inequalities on facilities: i F E(F ) z i 1 + E(F ) F CC(F (J)), J R (Cov F ) To see that these inequalities are valid we can rewrite inequalities (2) as i F (1 z i) F 1. The corresponding extended cover inequality is then (1 z i ) F 1. i F E(F ) Rewriting this inequality gives (Cov F ). The sets of inequalities (MC F ) and (Cov F ) do not contain each other as the following counterexamples show. In the example in Figure 3(a) a valid (Cov F ) inequality is z 1 +z 2 2, while the (MC F ) inequalities only ensure z 1 +z 2 +z 3 2. On the contrary, for the example given in Figure 3(b) the (Cov F ) inequalities are z 1 + z 2 1 and z 1 + z 3 1, but they are strictly dominated by the (MC F ) inequality z 1 + z 2 + z 3 2 that also implies z 2 + z 3 1. v 1 = 25 1 b 4 = 15 4 v 1 = 64 1 b 4 = 16 4 v 2 = 25 2 b 5 = 12 5 v 2 = 32 2 b 5 = 16 5 v 3 = 12 3 b 6 = 12 6 v 3 = 32 3 b 6 = 17 6 (a) Example 1 (b) Example 2 Fig. 3 Counterexamples for comparison of (MC F ) and (Cov F )

A cutting plane algorithm for the capacitated connected facility location problem 11 3.2.3 General representation of cover inequalities on facilities Consider now a general valid inequality of type z i p (3) i ˆF defined for a set ˆF F and p 1. For p = 1 we have the simple cover inequalities (2) (i.e., ˆF CC(F (J))) and, for p 2, inequalities of type (MC F ) and (Cov F ) belong to this family, i.e., we have ˆF F (J) {F E(F ) F CC(F (J))}, for J R. The following family of general cover inequalities on facilities is then also valid for our problem: z i 1 F F, F ˆF ˆF p + 1 (Cov gen ) i F It is not difficult to see that the latter inequalities are implied by (3). However, they are of particular interest when combined with cut set inequalities, as explained below. 3.3 Cut set cover inequalities This new family of valid inequalities combines cut set inequalities with the general cover inequalities for facilities of the form (Cov gen ). Inequalities (Cov gen ) state that at least one facility in F needs to be opened in a feasible solution. Consequently, for every subset of nodes W V containing all nodes in F, at least one ingoing arc needs to be installed. Let F denote the family of all subsets of facilities for which (Cov gen ) is valid: F = F (J) {F E(F ) F CC(F (J))} and let J R { p( ˆF 1 + E(F ), ˆF = F E(F ), F CC(F (J)) ) = p(j), ˆF = F (J) for all ˆF F. The following cut set cover inequalities are valid for CapConFL and not implied by any of the previously described sets of constraints: x ij 1 F W F, F ˆF ˆF p( ˆF ) + 1, ˆF F (CutCov ) ij δ (W ) Inequalities (Cut Cov ) are a generalization of the previously introduced cut set cover inequalities for the incremental ConFL studied in Arulselvan et al. [2]. Figure 4 illustrates inequalities (Cut Cov ) for two different subsets W and a cover inequality z 1 + z 2 1 of type (Cov F ). Figure 5 illustrates inequalities (Cut Cov ) for the minimum cardinality inequality z 1 + z 2 + z 3 2.

12 Stefan Gollowitzer et al. W 2 1 W 1 4 2 5 r 3 6 Fig. 4 Example for cut set cover inequalities (Cut Cov ) derived from an inequality (Cov F ). W 1 1 4 2 5 r 3 W 2 6 Fig. 5 Example for cut set cover inequalities (Cut Cov ) derived from an inequality (MC F ). 3.4 Cover inequalities for single cut sets The following set of valid inequalities generalizes the cover inequalities known for the capacitated network design problem studied in Chouman et al. [5]. Consider a (Cut SCF ) cut set inequality ij δ (W ) u ijx ij l F W d lz l defined by a cut set δ (W ) for W V \ {r}. Let F F W and d(f ) = l F d l. A set C δ (W ) is called a cover with respect to δ (W ) and F, if ij δ (W )\C u ij < d(f ) and a minimal cover if, in addition, u ij + u lk d(f ) (l, k) C. ij δ (W )\C Let MC(W, F ) denote the set of all minimal covers with respect to δ (W ) and F. Then the following cover inequalities on single cut sets are valid for the CapConFL: x ij 1 + (z l 1) C MC(W, F ). (Cov δ (W )) l F ij C Figure 6 illustrates inequalities (Cov δ (W )). Edge (b, d) is a cover with respect to W = {1, 2, 3, d, e, f} and F = {2, 3}. 4 Separation procedures In this section we describe the separation procedures used in our branch-andcut algorithm. We refer to the variable values of the current fractional solution by ( x, z).

A cutting plane algorithm for the capacitated connected facility location problem 13 d 1 = 5 r u = 30 b d u = 9 c u = 7 f e 1 2 3 d 2 = 7 d 3 = 12 Fig. 6 Illustration of cut set cover inequalities 4.1 Separation of inequalities (Cut SCF ) and (Cut Z ) Inequalities (Cut SCF ) can be separated in polynomial time (see also Ljubić et al. [29]). We define the support graph G = (V, A ) where V := V S t with an additional sink node t, A := A S A t and A t := {(i, t) i F, z i > 0}. We define capacities on arcs as u ij x ij for each arc ij A S and d i z i for each arc it A t. We calculate the minimum cut between r and t in G. Let δ (W ) denote the arcs of this cut. If δ (W ) A S and ij δ (W ) A S u ij x ij < i W F d i z i we have detected a violated inequality (Cut SCF ). Inequalities (Cut Z ) can be separated in similar fashion (see also Gollowitzer and Ljubić [19]). The support graph in this case is the bidirected core network (V S, A S ) with arc capacities set to x ij for each arc ij A S. A minimum cut in A S between r and l F with a weight of less than z l corresponds to a violated inequality (Cut Z ). 4.2 Separation of inequalities (Cut X ) For the separation of (Cut X ) inequalities we define a support graph G j for each j R. Thereby, G j = (V {j}, A S A j ) where A j = {(i, j) i F (j)}. Capacities on the arcs from A S A j are set to x ij. Each minimum cut in G j between r and j R whose weight is less than 1 corresponds to a violated inequality (Cut X ). If the number of customers is large, complete separation of inequalities (Cut X ) is very time-consuming. We therefore reduce the set of customers considered in the separation to a subset that still ensures that all violated inequalities are identified. A customer c 1 C is ignored if there exists another customer c 2 C such that F (c 2 ) F (c 1 ). If sets F (c i ) are identical for all c i C C only one customer in C is considered. 4.3 Separation of inequalities (EKS) For a fractional point ( x, z) and for each i F = {i F z i > 0}, the separation of (simple, non-extended) inequalities (EKS) is equivalent to solving

14 Stefan Gollowitzer et al. a knapsack problem which is described by the following integer program: min α = j R( z i x ij )s j s.t. b j s j > v i j R s j {0, 1} j R If α < z i a simple, non-extended cover inequality is violated. Instead of solving the separation problem exactly (e.g., using a dynamic programming procedure), we use a heuristic separation proposed by Kaparis and Letchford [24] that was shown to be effective in directly detecting extended knapsack cover inequalities. Adapted to our (EKS) inequalities, this procedure consists of the following steps, executed for each i F : 1. Sort the items in R(i) in non-decreasing order of ( z i x ij )/b j, and store them in a list L. Initialize the cover R as the empty set and initialize b = v i. 2. Remove an item from the head of the sorted list L. If its weight is larger than b, ignore it, otherwise insert it into R. If R is now a cover, go to step 4. 3. If L is empty, stop. Otherwise, return to step 2. 4. If the extended cover inequality corresponding to R is violated by ( x, z), output it. 5. Let k = arg max j R b j be the customer in R with the highest demand. Set b = b k and delete k from R. Return to step 2. In fact, we perform two variants of this algorithm. The one stated above and one where the items in R(i) are sorted in non-increasing order of x ij. 4.4 Separation of inequalities (MC F ) and (Cov F ) We consider subsets of facilities F F C := F 1 F 2, where F 1 := {F (k) k R} and F 2 := {F (k 1 ) F (k 2 ) F (k 1 ) F (k 2 ) / min( F (k 1 ), F (k 2 ) ) 0.5, k 1, k 2 R}, i.e., F 2 contains unions of F (k 1 ) and F (k 2 ) such that at least half the facilities of either F (k 1 ) or F (k 2 ) are common to both these sets. For each F we define the subset of customers to be considered in the separation of (MC F ) and (Cov F ) inequalities as J(F ) := {k R F (k ) F }. 4.4.1 Separation of inequalities (MC F ) We calculate p(j) for J and F (J) by solving the following bin-packing problem with assignment restrictions and non-uniform bin capacities: p(j) = min i F (J) t i

A cutting plane algorithm for the capacitated connected facility location problem 15 s.t. k R(i) i F (k) b k s ik v i t i s ik t i s ik = 1 s ik {0, 1} t i {0, 1} i F (J) k J, i F (k) k J k J, i F (k) i F (J) We consider F F C as candidate sets for F (J) and determine J = J(F ) as described in the previous paragraph. The values of p(j) are calculated for all such J by means of a general purpose solver during preprocessing. In the separation procedure we repeatedly check whether the current fractional solution violates any of the stored inequalities (MC F ). By doing so we consider at most F C R + R 2 inequalities of type (MC F ). 4.4.2 Separation of inequalities (Cov F ) Given J R and F (J), the separation of (simple, non-extended) covers on facilities (2) is equivalent to solving the following knapsack problem: min α = s.t. i F (J) i F (J) z i s i v i s i > i F (J) s i {0, 1} v i b(j) i F (J) If α < 1, an inequality (2) is violated. We consider F (J) for J = J(F ) and F F C. To find covers in CC(F (J)) we use the separation procedure described in Section 4.3 with the following modifications: The facilities in F (J) are ordered according to z i /v i in nondecreasing fashion and b is initialized with the maximum capacity of the facilities in F (J). 4.5 Separation of inequalities (Cut Cov ) In the separation of (Cut Cov ) we consider all inequalities of the form (3) that were found by the separation procedure for (Cov F ) and (MC F ). For the corresponding set of facilities ˆF and right-hand side p we randomly generate up to p sets F ˆF such that F = ˆF p + 1. We separate inequalities (Cut Cov ) by running a maximum flow algorithm on graph G defined as in Section 4.1, but with capacities of 1 on arcs (i, t) if i F and 0 if i F.

16 Stefan Gollowitzer et al. 4.6 Separation of inequalities (Cov δ (W )) Given a cut set W V \ {r} and the set of facilities contained in that cut set, F = W F, a violated cut set cover inequality is detected by solving the following integer program: min α = s.t. ij δ(w ) x ij s ij + l F (1 z l )t l ij δ (W ) u ij s ij + l F d l t l > ij δ (W ) u ij s ij {0, 1} (i, j) δ (W ) t l {0, 1} l F A (Cov δ (W )) inequality is violated if α < 1. We separate inequalities (Cov δ (W )) as follows: All cut sets W that are obtained during the separation of inequalities (Cut SCF ) are kept in a pool. We choose F = {l F W z l > 0.1}. Then we use the following heuristic procedure to find minimal covers C MC(W, F ), where F F : 1. Sort the items in δ (W ) and F in non-decreasing order of (1 z i )/d i and x ij /u ij and store them in a list L. Initialize the cover C and F as empty sets and initialize u = ij δ (W ) u ij. 2. Remove an item from the head of the sorted list L. (a) If it is an arc and its capacity is larger than u, ignore it, otherwise insert it into C. (b) If it is a facility insert it into F. If C is now a cover with respect to δ (W ) and F, go to step 4. 3. If L is empty, stop. Otherwise, return to step 2. 4. If the cover inequality corresponding to C MC(W, F ) is violated by ( x, z), output it. 5. Let (i, j ) = arg max ij C u ij be the arc in C with the highest capacity. Set u = u i j and delete (i, j ) from C. Return to step 2. 5 Computational results In this section we report the results of our computational experiments. They were performed on a desktop machine with an 8-core Intel Core i7 CPU at 2.80 GHz and 8 GB RAM. Each run was performed on a single core. We used the CPLEX [23] branch-and-cut framework, version 12.2. All cutting plane generation procedures provided by CPLEX are turned off unless stated explicitly. All heuristics provided by CPLEX are turned off. The other parameters are set to their default values.

A cutting plane algorithm for the capacitated connected facility location problem 17 5.1 Branch-and-cut framework The settings described in this section are the result of our preliminary testing. To reduce the number of constraints that need to be identified by our separation routines we add degree balance constraints and subtour elimination constraints for cycles of size two to our model: x ji x(δ + (i)) + z i (j, i) A S, i F (4a) x ji x(δ + (i)) (j, i) A S, i V S \ (F {r}) (4b) z i x(δ (i)) i F (4c) x ij x(δ (i)) (i, j) A S, i V S \ {r} (4d) x ij + x ji 1 (i, j) A S, i < j, i, j r (4e) In order to reduce the size of the linear programs solved throughout the process we relax constraints (1d) and add them only if they are violated. Separation procedures are called in the following order: (EKS) - (Cov F ) - (MC F ) - (1d) - (Cov δ (W )) - (Cut Cov ) - (Cut Z ) - (Cut X ) - (Cut SCF ). To prevent a tailing off effect of the separation procedures we stop separating valid inequalities if the lower bound has improved by less than 0.05% for the last 10 calls of the separation procedures. We apply this rule at each node of the branch-and-bound tree. Inequalities (Cov F ), (MC F ), (Cov δ (W )), (Cut Cov ) and (Cut SCF ) are only separated at the root node of the branch-and-bound tree. Inequalities (EKS) and (1d) are separated at every node, separation of (Cut X ) is done at every 10th node and separation of (Cut Z ) is done at every 100th node. To improve the computational efficiency of the separation procedures for cut set inequalities, we search for nested minimum cardinality cuts. To do so, all capacities in the respective separation graph are increased by some ɛ > 0. Thus, every detected violated cut contains the least possible number of arcs. We resolve the linear program after adding at most 30 violated inequalities of any class. Finally, we randomly choose the target nodes to search for violated cuts. 5.2 Instances We generated a set of realistic benchmark instances derived from real world data, which were provided by an Austrian telecommunications provider. The key figures for the instances we use are listed in Table 1. The real world data contain most of the information needed for complete CapConFL instances: The sets of facilities, Steiner nodes and edges of the core network; a set of customers with associated demands; a set of assignment arcs connecting customers and facilities, including their distance and an estimate of the bandwidth provided by the respective assignment arc; lengths of core edges and assignment arcs. These inputs define five graphs with different topologies

18 Stefan Gollowitzer et al. that will be denoted by A, B, C, D and E. To complete the instances with respect to the input required by CapConFL we applied the following steps: For each instance a minimum customer bandwidth is selected, assignment arcs that provide less than this bandwidth are removed. We chose 20, 25 and 30 MBit/s and denote this by 20, 25 and 30 in the instance label. At most 20 assignment arcs per customer are considered. Customers without assignment arcs are removed and facilities without assignment arcs are replaced by Steiner nodes. Steiner nodes with a degree of two and their adjacent edges are replaced by a single edge. A technology for each facility is randomly selected. For FTTB instances we consider the following combinations of capacity, demand and cost: (32, 4, 4000), (64, 5, 6000), (128, 7, 8000). For FTTC instances we choose between (64, 4, 13000), (128, 4, 16000) and (192, 4, 20000). Edge capacities are uniformly randomly selected from [0.7µ, 1.3µ], where µ is equal to the demand of the smallest set of facilities needed to feasibly assign the customers, given the facility capacities chosen before [8]. 5.3 Comparison against basic model and general purpose solver In the first part of our computational study we assess the influence of the cutting plane generation procedures built into CPLEX compared to the influence of the valid inequalities proposed in this work. To this end we ran our model with the following different settings: Basic is the cut set based model corresponding to SCF, i.e., the model consisting of constraints (1c)-(1g), (Cut SCF ) and (4a)-(4e). Basic+CPX is the basic model with all CPLEX cuts turned on. All VI is the basic model with all valid inequalities from Section 3 added. All VI+CPX is the basic model with all valid inequalities and CPLEX cuts turned on. In Table 1 we compare the LP gaps (g LP ) and time to solve the LP relaxation (t LP ) for these four models. We calculated the gaps as (UB LB)/UB, where UB is the best known integer solution found in all our tests and LB is the solution value of the LP relaxation of the respective model (after dynamic addition of cuts). In the last two lines we show the mean and median of the values in the respective column. The best LP gap of the four models is shown in bold. From the results in Table 1 we conclude that the model without valid inequalities provides a weak LP bound with an LP gap of 14.28% on average over the considered instance set. The cutting planes provided by CPLEX can reduce the LP gaps of the basic model by almost one half to an average of 7.23%. This average gap is still substantial compared to 1.31% obtained by the model that is strengthened by the valid inequalities proposed in this paper. Using CPLEX cuts in addition only improves the average gap to 1.13%.

A cutting plane algorithm for the capacitated connected facility location problem 19 Instance properties Basic Basic+CPX All VI All VI+CPX V E F R ES ER UB glp tlp glp tlp glp tlp glp tlp A20-FTTC 2405 12827 1504 805 1636 11191 3315257 12.98 10 7.67 45 3.56 33 3.56 36 A25-FTTC 2283 11349 1371 794 1525 9824 3717919 9.54 11 4.15 18 3.36 36 3.11 43 A30-FTTC 1949 7689 1053 694 1289 6400 3356273 8.62 5 4.74 12 3.45 29 2.77 38 B20-FTTC 2300 9695 1624 510 1824 7871 4601666 15.87 17 4.37 62 0.10 39 0.10 39 B25-FTTC 2265 9141 1580 507 1792 7349 5051559 14.42 13 5.37 29 0.15 36 0.07 41 B30-FTTC 1848 5610 1152 450 1432 4178 6030938 9.18 8 4.06 16 0.10 17 0.02 18 C20-FTTC 4742 21388 3237 1160 3720 17668 8897504 21.06 149 11.12 409 0.45 141 0.36 152 C25-FTTC 4498 17477 2961 1134 3498 13979 9734526 16.29 55 9.91 228 0.57 134 0.53 145 C30-FTTC 3269 9303 1748 840 2557 6746 9332930 11.14 47 4.73 75 0.82 35 0.70 58 D20-FTTC 4042 24923 2241 1463 2661 22262 7962207 13.18 50 8.85 156 1.93 63 1.90 79 D25-FTTC 3925 21403 2106 1449 2557 18846 9490801 8.86 38 4.31 76 1.32 62 1.01 74 D30-FTTC 3407 11539 1588 1213 2274 9265 9971466 4.03 12 2.66 30 0.72 61 0.45 73 E20-FTTC 3426 18045 2143 1038 2492 15553 4858219 23.95 101 15.81 302 1.06 61 0.86 93 E25-FTTC 3290 12574 1965 1023 2369 10205 7419660 13.21 58 5.13 114 0.54 51 0.34 74 E30-FTTC 2149 4429 750 497 1747 2682 6082850 6.93 7 2.54 9 0.00 21 0.00 21 A20-FTTB 2405 12827 1504 805 1636 11191 2506131 16.00 10 9.56 37 3.23 40 2.42 48 A25-FTTB 2267 11284 1370 778 1525 9759 2741939 12.64 8 7.27 33 3.26 38 2.81 47 A30-FTTB 1949 7689 1053 694 1289 6400 2406722 10.01 4 3.75 17 2.73 26 2.14 37 B20-FTTB 2300 9695 1624 510 1824 7871 3781915 19.99 13 7.83 66 0.26 40 0.21 49 B25-FTTB 2265 9141 1580 507 1792 7349 4146838 18.22 11 5.77 31 0.17 37 0.07 37 B30-FTTB 1848 5610 1152 450 1432 4178 4797642 12.12 10 3.75 18 0.06 19 0.05 20 C20-FTTB 4742 21388 3237 1160 3720 17668 7337282 26.20 176 14.11 397 0.62 158 0.56 171 C25-FTTB 4498 17477 2961 1134 3498 13979 7836463 20.75 66 10.93 229 0.78 120 0.69 159 C30-FTTB 3269 9303 1748 840 2557 6746 7332424 14.18 44 6.85 63 0.89 56 0.79 61 D20-FTTB 4042 24923 2241 1463 2661 22262 7056336 13.94 46 9.30 131 1.06 57 1.02 68 D25-FTTB 3925 21403 2106 1449 2557 18846 8315107 10.26 39 6.22 78 1.44 65 1.20 71 D30-FTTB 3407 11539 1588 1213 2274 9265 8210407 4.67 11 3.05 28 0.56 50 0.41 60 E20-FTTB 3426 18045 2143 1038 2492 15553 3956572 29.98 188 19.18 413 1.38 64 1.23 88 E25-FTTB 3290 12574 1965 1023 2369 10205 5885571 19.53 99 9.27 170 3.07 65 2.89 60 E30-FTTB 2149 4429 750 497 1747 2682 4651295 10.53 5 4.66 9 1.55 24 1.48 23 mean 14.28 44 7.23 110 1.31 56 1.13 66 median 13.19 15 6.00 62 0.86 45 0.75 59 Table 1 Comparison of LP relaxation lower bounds with and without CPLEX cuts

20 Stefan Gollowitzer et al. 5.4 Influence of different sets of valid inequalities In the second part of our computational study we assess the influence of the different sets of valid inequalities proposed in this work. We compare five different settings that differ by the sets of valid inequalities considered. For each setting we add a subset of valid inequalities to the basic model described above. Settings (Cut Z ), (Cut X ) and (Cut Z )+(Cut X ) are self-explaining. Setting All VI is defined as above and setting Most VI uses inequalities (Cut Z ), (Cut X ), (EKS), (Cov F ) and (MC F ). For each of these settings, Table 2 shows the gap of the linear programming relaxation, g LP, after the dynamic addition of cuts and calculated as in Table 1, the time needed to solve linear programming relaxation, t LP, and the number of cutting planes added, Cuts. In the last two lines we show the mean and median of the values in the respective column. The best LP gap of all models is shown in bold. We would like to point out several interesting aspects. The LP gaps of setting (Cut Z ) are substantially larger than the ones of all other settings. Surprisingly the same does not hold for setting (Cut X ), which on average gives even stronger LP bounds than setting (Cut Z )+(Cut X ). We trace the difference between the gaps of (Cut X ) and (Cut Z )+(Cut X ) to the criteria we used to prevent a tailing off effect during separation. A comparison of the running times shows that separating valid inequalities with a different structure improves the overall running time of the LP relaxation. Approaches (Cut Z ) and (Cut X ) need 110 and 90 seconds on average, respectively, whereas approach (Cut Z )+(Cut X ) only takes 46 seconds to compute approximately the same lower bounds as (Cut X ). The separation routines in approaches Most VI and All VI require an additional 10 and 12 seconds on average. Thereby, the average LP gaps are improved from 1.44% ((Cut Z )+(Cut X )) to 1.31% (Most VI and All VI ). However, All VI does not improve upon Most VI significantly. There is a notable difference in the numbers of valid inequalities that were detected during the LP relaxations of the different settings. By far the highest number of inequalities is found by setting (Cut Z ), even though the obtained LP bound is comparably weak. This is consistent with the long running time of the LP relaxation of setting (Cut Z ). Rather surprising is the fact, that setting All VI obtains the same LP bound as setting Most VI for 28 out of 30 settings but the number of valid inequalities found by All VI is smaller for 22 and larger for only 2 instances. Table 3 shows the respective gap of the five different settings after 3, 10, 30 and 60 minutes. For these results we calculate the gaps as (UB t LB t )/UB t where UB t is the best integer solution found by the respective setting after t minutes and LB t is the lower bound after t minutes. For each instance and running time the smallest gap of all five settings is indicated in bold. If no integer solution is available after t minutes we indicate this by a dash in the respective column. For each setting and time t the last three lines of the table indicate the mean and median of gaps over the instance set and how often the respective approach gives the smallest gap of all settings.

A cutting plane algorithm for the capacitated connected facility location problem 21 (CutZ) (CutX) (CutZ)+(CutX) Most VI All VI glp tlp Cuts glp tlp Cuts glp tlp Cuts glp tlp Cuts glp tlp Cuts A20-FTTC 8.56 32 5711 3.57 48 3879 3.58 27 4168 3.56 31 4776 3.56 33 4523 A25-FTTC 4.48 20 4623 3.50 94 2984 3.51 31 3101 3.38 32 3528 3.36 36 3446 A30-FTTC 7.12 17 3565 4.05 47 2987 4.05 24 3144 3.45 27 3855 3.45 29 3660 B20-FTTC 11.85 105 7323 0.15 39 4459 0.15 26 5100 0.10 39 5140 0.10 39 5140 B25-FTTC 10.40 26 5669 0.18 50 4442 0.18 33 4815 0.15 35 4908 0.15 36 4908 B30-FTTC 5.03 17 3447 0.12 24 2382 0.12 16 2661 0.10 17 2728 0.10 17 2728 C20-FTTC 12.98 351 14191 0.54 293 9042 0.54 144 9813 0.45 142 10114 0.45 141 9831 C25-FTTC 8.32 265 11612 0.62 215 6947 0.62 122 7298 0.56 122 8107 0.57 134 7710 C30-FTTC 4.91 62 4514 0.98 79 3290 0.98 32 3809 0.82 35 3873 0.82 35 3799 D20-FTTC 7.79 232 13921 2.00 97 7561 2.00 51 8193 1.93 62 8751 1.93 63 8124 D25-FTTC 3.68 62 8784 1.33 68 5901 1.34 39 6056 1.32 62 6454 1.32 62 6529 D30-FTTC 1.28 50 6150 0.82 109 4578 0.82 55 4737 0.72 58 5001 0.72 61 4974 E20-FTTC 13.69 142 12891 1.10 140 8633 1.10 58 9566 1.06 61 10021 1.06 61 10021 E25-FTTC 4.17 50 8515 0.57 62 6046 0.58 35 6303 0.54 50 6656 0.54 51 6656 E30-FTTC 2.29 12 2222 0.01 20 1682 0.01 14 1795 0.00 22 1960 0.00 21 1849 A20-FTTB 10.46 25 5426 3.46 67 3681 3.58 38 3640 3.23 40 4880 3.23 40 4394 A25-FTTB 5.09 20 3857 3.53 78 2740 3.57 29 2851 3.26 32 3789 3.26 38 3457 A30-FTTB 8.25 14 2755 3.59 26 2036 3.60 13 2133 2.73 22 3074 2.73 26 2832 B20-FTTB 15.82 74 7895 0.36 54 4497 0.36 34 4977 0.26 38 5365 0.26 40 4944 B25-FTTB 13.71 45 6044 0.26 50 4503 0.26 32 4948 0.17 37 5051 0.17 37 4913 B30-FTTB 7.20 16 3478 0.11 25 2387 0.11 18 2661 0.06 22 2601 0.06 19 2664 C20-FTTB 16.75 657 15243 0.71 298 8207 0.71 148 8980 0.62 156 9785 0.62 158 9475 C25-FTTB 11.26 306 12195 0.88 93 6765 0.88 67 7266 0.78 118 7793 0.78 120 7364 C30-FTTB 6.73 84 4785 1.10 93 3380 1.10 50 3693 0.90 57 3942 0.89 56 3756 D20-FTTB 8.39 210 12375 1.10 65 5535 1.10 30 6459 1.06 59 7591 1.06 57 6476 D25-FTTB 4.56 75 8617 1.46 154 4824 1.46 50 4950 1.44 59 6104 1.44 65 5565 D30-FTTB 1.34 28 6054 0.67 43 4030 0.67 24 4287 0.56 46 4543 0.56 50 4308 E20-FTTB 18.30 213 12824 1.39 139 7394 1.39 68 7869 1.38 74 8747 1.38 64 8379 E25-FTTB 8.41 82 8547 3.09 113 5820 3.12 61 6010 3.07 50 6699 3.07 65 6443 E30-FTTB 5.36 20 2288 1.68 10 1617 1.68 13 1811 1.55 24 1937 1.55 24 1937 mean 8.27 110 7517 1.43 90 4740 1.44 46 5103 1.31 54 5592 1.31 56 5360 median 8.02 56 6102 1.04 67 4478 1.04 33 4881 0.86 43 5026 0.86 45 4928 Table 2 Comparison of LP relaxation lower bounds with different sets of valid inequalities turned on

22 Stefan Gollowitzer et al. (CutZ) (CutX) (CutZ)+(CutX) Most VI All VI g3 g10 g30 g60 g3 g10 g30 g60 g3 g10 g30 g60 g3 g10 g30 g60 g3 g10 g30 g60 A20-FTTC 7.28 5.44 4.97 4.12 - - 4.76 4.76 4.25 4.22 4.03 3.45 4.34 4.22 4.21 4.18 3.68 3.66 3.62 3.61 A25-FTTC 5.54 5.26 5.04 4.96 - - 4.99 4.96 5.88 5.82 5.77 5.74 4.44 4.37 4.32 4.29 3.74 3.66 3.61 3.58 A30-FTTC 6.65 5.82 5.31 5.17-5.81 5.66 5.52 5.72 5.55 4.37 3.93 4.87 3.61 3.31 3.02 4.93 3.58 3.40 2.86 B20-FTTC 11.59 5.55 3.44 2.13 0.07 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 B25-FTTC 7.58 4.65 2.25 1.20 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 B30-FTTC 1.48 0.69 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 C20-FTTC - 15.91 6.21 5.03 - - 1.94 1.48-1.83 1.79 0.61-0.82 0.59 0.32-1.00 0.52 0.46 C25-FTTC - 5.85 4.09 3.72-1.03 0.99 0.90 0.99 0.93 0.88 0.85 1.06 0.86 0.70 0.66 0.83 0.73 0.65 0.62 C30-FTTC 5.22 3.37 3.00 2.68 1.47 1.42 1.33 1.24 1.34 1.19 1.10 1.01 0.95 0.82 0.30 0.05 0.91 0.73 0.13 0.00 D20-FTTC - - 3.34 2.83 - - 2.59 2.59-2.49 2.46 2.42-2.53 2.00 1.90-2.15 2.10 2.07 D25-FTTC 4.14 2.37 2.07 1.77 - - 1.99 1.80 2.39 1.61 1.24 1.18-1.62 1.18 1.06-1.60 1.46 1.13 D30-FTTC 1.53 1.19 0.96 0.89-1.31 1.23 0.96 1.23 0.69 0.55 0.53 1.17 1.04 0.87 0.59 1.20 0.91 0.60 0.58 E20-FTTC - 8.32 5.50 4.25-1.98 1.77 1.66 1.79 1.43 1.22 1.16 1.73 1.21 0.85 0.80 1.73 1.21 0.85 0.80 E25-FTTC 4.37 2.56 2.24 1.85 1.99 1.82 0.70 0.67 0.70 0.55 0.42 0.32 0.73 0.59 0.46 0.35 0.73 0.59 0.46 0.35 E30-FTTC 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 A20-FTTB 8.31 7.29 5.56 5.21 - - 5.07 5.05 5.73 5.68 5.62 5.49-3.17 3.06 2.96-2.70 2.53 2.43 A25-FTTB 6.90 6.58 6.27 4.47-3.52 3.45 3.37 4.82 4.73 4.59 4.46 3.50 3.26 3.01 2.97 3.38 3.19 3.05 3.02 A30-FTTB 6.85 5.22 4.86 4.65-6.13 5.98 5.95 5.06 4.88 4.78 4.73 2.94 2.66 2.29 2.10 2.84 2.63 2.42 2.34 B20-FTTB 16.17 11.15 6.52 4.58 0.65 0.60 0.55 0.43 0.44 0.37 0.25 0.23 0.30 0.27 0.16 0.14 0.32 0.22 0.18 0.16 B25-FTTB 10.99 7.87 5.22 4.11 0.58 0.47 0.38 0.34 0.23 0.10 0.05 0.03 0.04 0.00 0.00 0.00 0.06 0.02 0.00 0.00 B30-FTTB 3.57 1.41 0.00 0.00 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 C20-FTTB - - 11.43 8.69 - - 1.32 1.31-1.13 0.88 0.85-0.76 0.72 0.58-0.80 0.78 0.72 C25-FTTB - - 5.43 4.68 - - 1.84 1.19-1.32 1.02 1.00-0.89 0.69 0.65-0.85 0.61 0.46 C30-FTTB 7.32 2.79 1.16 0.81-0.69 0.66 0.62 0.59 0.43 0.37 0.34 0.42 0.35 0.27 0.25 0.33 0.25 0.21 0.16 D20-FTTB - 4.68 2.73 2.25 - - - 1.41-1.21 1.19 1.17-1.39 1.17 1.14-1.19 1.07 1.04 D25-FTTB - 2.96 2.10 1.85 - - - 2.29-1.77 1.72 1.70-1.93 1.49 1.45-2.28 1.31 1.29 D30-FTTB 1.40 1.12 0.92 0.80-0.83 0.70 0.62 0.87 0.58 0.43 0.41 0.64 0.51 0.36 0.33 0.54 0.47 0.40 0.33 E20-FTTB - 10.18 6.09 4.47 - - 1.66 1.63 2.33 1.52 1.36 1.35 2.57 1.79 1.34 1.22 1.61 1.54 1.33 1.31 E25-FTTB - 3.92 2.39 2.13-1.68 0.87 0.71 0.77 0.51 0.44 0.41 1.41 0.62 0.45 0.28 0.91 0.57 0.50 0.44 E30-FTTB 2.18 1.15 0.12 0.10 0.40 0.20 0.16 0.15 0.14 0.12 0.10 0.08 0.14 0.04 0.00 0.00 0.14 0.04 0.00 0.00 mean - - 3.64 2.98 - - - 1.72-1.69 1.55 1.45-1.31 1.13 1.04-1.22 1.06 0.99 median - - 3.39 2.75 - - - 1.21-1.16 0.95 0.85-0.84 0.70 0.58-0.83 0.61 0.52 best 1 1 3 3 3 5 5 5 10 10 8 8 11 10 16 18 16 22 19 18 Table 3 Gaps to model upper bound